Beschreibung
Master's Thesis from the year 2018 in the subject Engineering - Robotics, National Institute of Technology, Rourkela, language: English, abstract: Indoor navigation is a challenging task due to the absence of Global Positioning System(GPS). This project removes the need for GPS in systems by combining Inertial Navigation Systems (INS) and Visual Navigation Systems (VNS), with the help of machine learning with Artificial and Convolutional Neural Networks.In GPS denied environments a highly accurate INS is necessary, it must also be coupled with another system to bound the continious drift error that is present in INS, for which VNS is employed. The system was implemented using a ground robot to collect ground truth data, which were used as datasets to train a filter that increases the accuracy of the INS. The accuracy of the INS has been proven on the hardware platfrom over multiple datasets. Eventually Visual Navigation data can also be fed into the same system, which for now is implemented in simulation, as an independent system. A software and hardware framework have been developed that can be used in the future for further developments. The project also optimizes visual navigation for use on low power hardware with hardware acceleration for maximized speed. A low cost and scalable indoor navigation system is developed for indoor navigation, which can also be further extended to Autonomous Underwater Vehicles (AUV) in 3D space.
